Smoothing and tying for Korean flexible vocabulary isolated word recognition

For large vocabulary recognition system, as well as for flexible vocabulary applications using hidden Markov model(HMM), parameter smoothing and tying have been used to increase the reliability of models. This paper describes bottom-up and topdown clustering techniques for state level tying. This paper also describes a method of applying parameter smoothing to the clustered states and covariance matrix of semicontinuous hidden Markov model(SCHMM). We present a new parameter smoothing method and apply it to the distribution of discrete hidden Markov model(DHMM) in the training procedure. A new model composition method for unseen triphone modeling in bottom-up clustering is also proposed and compared with traditional context-independent model backing-off method.